2 * An async IO implementation for Linux
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
9 * See ../COPYING for licensing terms.
11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
55 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
65 struct io_event io_events[0];
66 }; /* 128 bytes + ring size */
68 #define AIO_RING_PAGES 8
73 struct kioctx *table[];
77 unsigned reqs_available;
81 struct percpu_ref users;
84 struct percpu_ref reqs;
86 unsigned long user_id;
88 struct __percpu kioctx_cpu *cpu;
91 * For percpu reqs_available, number of slots we move to/from global
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
99 * The real limit is nr_events - 1, which will be larger (see
104 /* Size of ringbuffer, in units of struct io_event */
107 unsigned long mmap_base;
108 unsigned long mmap_size;
110 struct page **ring_pages;
113 struct work_struct free_work;
116 * signals when all in-flight requests are done
118 struct completion *requests_done;
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
127 * We batch accesses to it with a percpu version.
129 atomic_t reqs_available;
130 } ____cacheline_aligned_in_smp;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
138 struct mutex ring_lock;
139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
144 unsigned completed_events;
145 spinlock_t completion_lock;
146 } ____cacheline_aligned_in_smp;
148 struct page *internal_pages[AIO_RING_PAGES];
149 struct file *aio_ring_file;
155 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
156 * cancelled or completed (this makes a certain amount of sense because
157 * successful cancellation - io_cancel() - does deliver the completion to
160 * And since most things don't implement kiocb cancellation and we'd really like
161 * kiocb completion to be lockless when possible, we use ki_cancel to
162 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
163 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
165 #define KIOCB_CANCELLED ((void *) (~0ULL))
170 struct kioctx *ki_ctx;
171 kiocb_cancel_fn *ki_cancel;
173 struct iocb __user *ki_user_iocb; /* user's aiocb */
174 __u64 ki_user_data; /* user's data for completion */
176 struct list_head ki_list; /* the aio core uses this
177 * for cancellation */
180 * If the aio_resfd field of the userspace iocb is not zero,
181 * this is the underlying eventfd context to deliver events to.
183 struct eventfd_ctx *ki_eventfd;
186 /*------ sysctl variables----*/
187 static DEFINE_SPINLOCK(aio_nr_lock);
188 unsigned long aio_nr; /* current system wide number of aio requests */
189 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
190 /*----end sysctl variables---*/
192 static struct kmem_cache *kiocb_cachep;
193 static struct kmem_cache *kioctx_cachep;
195 static struct vfsmount *aio_mnt;
197 static const struct file_operations aio_ring_fops;
198 static const struct address_space_operations aio_ctx_aops;
200 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
202 struct qstr this = QSTR_INIT("[aio]", 5);
205 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
207 return ERR_CAST(inode);
209 inode->i_mapping->a_ops = &aio_ctx_aops;
210 inode->i_mapping->private_data = ctx;
211 inode->i_size = PAGE_SIZE * nr_pages;
213 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
216 return ERR_PTR(-ENOMEM);
218 path.mnt = mntget(aio_mnt);
220 d_instantiate(path.dentry, inode);
221 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227 file->f_flags = O_RDWR;
231 static struct dentry *aio_mount(struct file_system_type *fs_type,
232 int flags, const char *dev_name, void *data)
234 static const struct dentry_operations ops = {
235 .d_dname = simple_dname,
237 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
241 * Creates the slab caches used by the aio routines, panic on
242 * failure as this is done early during the boot sequence.
244 static int __init aio_setup(void)
246 static struct file_system_type aio_fs = {
249 .kill_sb = kill_anon_super,
251 aio_mnt = kern_mount(&aio_fs);
253 panic("Failed to create aio fs mount.");
255 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
256 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
258 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
262 __initcall(aio_setup);
264 static void put_aio_ring_file(struct kioctx *ctx)
266 struct file *aio_ring_file = ctx->aio_ring_file;
268 truncate_setsize(aio_ring_file->f_inode, 0);
270 /* Prevent further access to the kioctx from migratepages */
271 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
272 aio_ring_file->f_inode->i_mapping->private_data = NULL;
273 ctx->aio_ring_file = NULL;
274 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
280 static void aio_free_ring(struct kioctx *ctx)
284 /* Disconnect the kiotx from the ring file. This prevents future
285 * accesses to the kioctx from page migration.
287 put_aio_ring_file(ctx);
289 for (i = 0; i < ctx->nr_pages; i++) {
291 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
292 page_count(ctx->ring_pages[i]));
293 page = ctx->ring_pages[i];
296 ctx->ring_pages[i] = NULL;
300 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
301 kfree(ctx->ring_pages);
302 ctx->ring_pages = NULL;
306 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
308 vma->vm_flags |= VM_DONTEXPAND;
309 vma->vm_ops = &generic_file_vm_ops;
313 static int aio_ring_remap(struct file *file, struct vm_area_struct *vma)
315 struct mm_struct *mm = vma->vm_mm;
316 struct kioctx_table *table;
317 int i, res = -EINVAL;
319 spin_lock(&mm->ioctx_lock);
321 table = rcu_dereference(mm->ioctx_table);
322 for (i = 0; i < table->nr; i++) {
325 ctx = table->table[i];
326 if (ctx && ctx->aio_ring_file == file) {
327 if (!atomic_read(&ctx->dead)) {
328 ctx->user_id = ctx->mmap_base = vma->vm_start;
336 spin_unlock(&mm->ioctx_lock);
340 static const struct file_operations aio_ring_fops = {
341 .mmap = aio_ring_mmap,
342 .mremap = aio_ring_remap,
345 #if IS_ENABLED(CONFIG_MIGRATION)
346 static int aio_migratepage(struct address_space *mapping, struct page *new,
347 struct page *old, enum migrate_mode mode)
356 /* mapping->private_lock here protects against the kioctx teardown. */
357 spin_lock(&mapping->private_lock);
358 ctx = mapping->private_data;
364 /* The ring_lock mutex. The prevents aio_read_events() from writing
365 * to the ring's head, and prevents page migration from mucking in
366 * a partially initialized kiotx.
368 if (!mutex_trylock(&ctx->ring_lock)) {
374 if (idx < (pgoff_t)ctx->nr_pages) {
375 /* Make sure the old page hasn't already been changed */
376 if (ctx->ring_pages[idx] != old)
384 /* Writeback must be complete */
385 BUG_ON(PageWriteback(old));
388 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
389 if (rc != MIGRATEPAGE_SUCCESS) {
394 /* Take completion_lock to prevent other writes to the ring buffer
395 * while the old page is copied to the new. This prevents new
396 * events from being lost.
398 spin_lock_irqsave(&ctx->completion_lock, flags);
399 migrate_page_copy(new, old);
400 BUG_ON(ctx->ring_pages[idx] != old);
401 ctx->ring_pages[idx] = new;
402 spin_unlock_irqrestore(&ctx->completion_lock, flags);
404 /* The old page is no longer accessible. */
408 mutex_unlock(&ctx->ring_lock);
410 spin_unlock(&mapping->private_lock);
415 static const struct address_space_operations aio_ctx_aops = {
416 .set_page_dirty = __set_page_dirty_no_writeback,
417 #if IS_ENABLED(CONFIG_MIGRATION)
418 .migratepage = aio_migratepage,
422 static int aio_setup_ring(struct kioctx *ctx)
424 struct aio_ring *ring;
425 unsigned nr_events = ctx->max_reqs;
426 struct mm_struct *mm = current->mm;
427 unsigned long size, unused;
432 /* Compensate for the ring buffer's head/tail overlap entry */
433 nr_events += 2; /* 1 is required, 2 for good luck */
435 size = sizeof(struct aio_ring);
436 size += sizeof(struct io_event) * nr_events;
438 nr_pages = PFN_UP(size);
442 file = aio_private_file(ctx, nr_pages);
444 ctx->aio_ring_file = NULL;
448 ctx->aio_ring_file = file;
449 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
450 / sizeof(struct io_event);
452 ctx->ring_pages = ctx->internal_pages;
453 if (nr_pages > AIO_RING_PAGES) {
454 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
456 if (!ctx->ring_pages) {
457 put_aio_ring_file(ctx);
462 for (i = 0; i < nr_pages; i++) {
464 page = find_or_create_page(file->f_inode->i_mapping,
465 i, GFP_HIGHUSER | __GFP_ZERO);
468 pr_debug("pid(%d) page[%d]->count=%d\n",
469 current->pid, i, page_count(page));
470 SetPageUptodate(page);
473 ctx->ring_pages[i] = page;
477 if (unlikely(i != nr_pages)) {
482 ctx->mmap_size = nr_pages * PAGE_SIZE;
483 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
485 down_write(&mm->mmap_sem);
486 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
487 PROT_READ | PROT_WRITE,
488 MAP_SHARED, 0, &unused);
489 up_write(&mm->mmap_sem);
490 if (IS_ERR((void *)ctx->mmap_base)) {
496 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
498 ctx->user_id = ctx->mmap_base;
499 ctx->nr_events = nr_events; /* trusted copy */
501 ring = kmap_atomic(ctx->ring_pages[0]);
502 ring->nr = nr_events; /* user copy */
504 ring->head = ring->tail = 0;
505 ring->magic = AIO_RING_MAGIC;
506 ring->compat_features = AIO_RING_COMPAT_FEATURES;
507 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
508 ring->header_length = sizeof(struct aio_ring);
510 flush_dcache_page(ctx->ring_pages[0]);
515 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
516 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
517 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
519 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
521 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
522 struct kioctx *ctx = req->ki_ctx;
525 spin_lock_irqsave(&ctx->ctx_lock, flags);
527 if (!req->ki_list.next)
528 list_add(&req->ki_list, &ctx->active_reqs);
530 req->ki_cancel = cancel;
532 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
534 EXPORT_SYMBOL(kiocb_set_cancel_fn);
536 static int kiocb_cancel(struct aio_kiocb *kiocb)
538 kiocb_cancel_fn *old, *cancel;
541 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
542 * actually has a cancel function, hence the cmpxchg()
545 cancel = ACCESS_ONCE(kiocb->ki_cancel);
547 if (!cancel || cancel == KIOCB_CANCELLED)
551 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
552 } while (cancel != old);
554 return cancel(&kiocb->common);
557 static void free_ioctx(struct work_struct *work)
559 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
561 pr_debug("freeing %p\n", ctx);
564 free_percpu(ctx->cpu);
565 percpu_ref_exit(&ctx->reqs);
566 percpu_ref_exit(&ctx->users);
567 kmem_cache_free(kioctx_cachep, ctx);
570 static void free_ioctx_reqs(struct percpu_ref *ref)
572 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
574 /* At this point we know that there are no any in-flight requests */
575 if (ctx->requests_done)
576 complete(ctx->requests_done);
578 INIT_WORK(&ctx->free_work, free_ioctx);
579 schedule_work(&ctx->free_work);
583 * When this function runs, the kioctx has been removed from the "hash table"
584 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
585 * now it's safe to cancel any that need to be.
587 static void free_ioctx_users(struct percpu_ref *ref)
589 struct kioctx *ctx = container_of(ref, struct kioctx, users);
590 struct aio_kiocb *req;
592 spin_lock_irq(&ctx->ctx_lock);
594 while (!list_empty(&ctx->active_reqs)) {
595 req = list_first_entry(&ctx->active_reqs,
596 struct aio_kiocb, ki_list);
598 list_del_init(&req->ki_list);
602 spin_unlock_irq(&ctx->ctx_lock);
604 percpu_ref_kill(&ctx->reqs);
605 percpu_ref_put(&ctx->reqs);
608 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
611 struct kioctx_table *table, *old;
612 struct aio_ring *ring;
614 spin_lock(&mm->ioctx_lock);
615 table = rcu_dereference_raw(mm->ioctx_table);
619 for (i = 0; i < table->nr; i++)
620 if (!table->table[i]) {
622 table->table[i] = ctx;
623 spin_unlock(&mm->ioctx_lock);
625 /* While kioctx setup is in progress,
626 * we are protected from page migration
627 * changes ring_pages by ->ring_lock.
629 ring = kmap_atomic(ctx->ring_pages[0]);
635 new_nr = (table ? table->nr : 1) * 4;
636 spin_unlock(&mm->ioctx_lock);
638 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
645 spin_lock(&mm->ioctx_lock);
646 old = rcu_dereference_raw(mm->ioctx_table);
649 rcu_assign_pointer(mm->ioctx_table, table);
650 } else if (table->nr > old->nr) {
651 memcpy(table->table, old->table,
652 old->nr * sizeof(struct kioctx *));
654 rcu_assign_pointer(mm->ioctx_table, table);
663 static void aio_nr_sub(unsigned nr)
665 spin_lock(&aio_nr_lock);
666 if (WARN_ON(aio_nr - nr > aio_nr))
670 spin_unlock(&aio_nr_lock);
674 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
676 static struct kioctx *ioctx_alloc(unsigned nr_events)
678 struct mm_struct *mm = current->mm;
683 * We keep track of the number of available ringbuffer slots, to prevent
684 * overflow (reqs_available), and we also use percpu counters for this.
686 * So since up to half the slots might be on other cpu's percpu counters
687 * and unavailable, double nr_events so userspace sees what they
688 * expected: additionally, we move req_batch slots to/from percpu
689 * counters at a time, so make sure that isn't 0:
691 nr_events = max(nr_events, num_possible_cpus() * 4);
694 /* Prevent overflows */
695 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
696 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
697 pr_debug("ENOMEM: nr_events too high\n");
698 return ERR_PTR(-EINVAL);
701 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
702 return ERR_PTR(-EAGAIN);
704 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
706 return ERR_PTR(-ENOMEM);
708 ctx->max_reqs = nr_events;
710 spin_lock_init(&ctx->ctx_lock);
711 spin_lock_init(&ctx->completion_lock);
712 mutex_init(&ctx->ring_lock);
713 /* Protect against page migration throughout kiotx setup by keeping
714 * the ring_lock mutex held until setup is complete. */
715 mutex_lock(&ctx->ring_lock);
716 init_waitqueue_head(&ctx->wait);
718 INIT_LIST_HEAD(&ctx->active_reqs);
720 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
723 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
726 ctx->cpu = alloc_percpu(struct kioctx_cpu);
730 err = aio_setup_ring(ctx);
734 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
735 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
736 if (ctx->req_batch < 1)
739 /* limit the number of system wide aios */
740 spin_lock(&aio_nr_lock);
741 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
742 aio_nr + nr_events < aio_nr) {
743 spin_unlock(&aio_nr_lock);
747 aio_nr += ctx->max_reqs;
748 spin_unlock(&aio_nr_lock);
750 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
751 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
753 err = ioctx_add_table(ctx, mm);
757 /* Release the ring_lock mutex now that all setup is complete. */
758 mutex_unlock(&ctx->ring_lock);
760 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
761 ctx, ctx->user_id, mm, ctx->nr_events);
765 aio_nr_sub(ctx->max_reqs);
767 atomic_set(&ctx->dead, 1);
769 vm_munmap(ctx->mmap_base, ctx->mmap_size);
772 mutex_unlock(&ctx->ring_lock);
773 free_percpu(ctx->cpu);
774 percpu_ref_exit(&ctx->reqs);
775 percpu_ref_exit(&ctx->users);
776 kmem_cache_free(kioctx_cachep, ctx);
777 pr_debug("error allocating ioctx %d\n", err);
782 * Cancels all outstanding aio requests on an aio context. Used
783 * when the processes owning a context have all exited to encourage
784 * the rapid destruction of the kioctx.
786 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
787 struct completion *requests_done)
789 struct kioctx_table *table;
791 spin_lock(&mm->ioctx_lock);
792 if (atomic_xchg(&ctx->dead, 1)) {
793 spin_unlock(&mm->ioctx_lock);
797 table = rcu_dereference_raw(mm->ioctx_table);
798 WARN_ON(ctx != table->table[ctx->id]);
799 table->table[ctx->id] = NULL;
800 spin_unlock(&mm->ioctx_lock);
802 /* percpu_ref_kill() will do the necessary call_rcu() */
803 wake_up_all(&ctx->wait);
806 * It'd be more correct to do this in free_ioctx(), after all
807 * the outstanding kiocbs have finished - but by then io_destroy
808 * has already returned, so io_setup() could potentially return
809 * -EAGAIN with no ioctxs actually in use (as far as userspace
812 aio_nr_sub(ctx->max_reqs);
815 vm_munmap(ctx->mmap_base, ctx->mmap_size);
817 ctx->requests_done = requests_done;
818 percpu_ref_kill(&ctx->users);
823 * exit_aio: called when the last user of mm goes away. At this point, there is
824 * no way for any new requests to be submited or any of the io_* syscalls to be
825 * called on the context.
827 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
830 void exit_aio(struct mm_struct *mm)
832 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
838 for (i = 0; i < table->nr; ++i) {
839 struct kioctx *ctx = table->table[i];
840 struct completion requests_done =
841 COMPLETION_INITIALIZER_ONSTACK(requests_done);
846 * We don't need to bother with munmap() here - exit_mmap(mm)
847 * is coming and it'll unmap everything. And we simply can't,
848 * this is not necessarily our ->mm.
849 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
850 * that it needs to unmap the area, just set it to 0.
853 kill_ioctx(mm, ctx, &requests_done);
855 /* Wait until all IO for the context are done. */
856 wait_for_completion(&requests_done);
859 RCU_INIT_POINTER(mm->ioctx_table, NULL);
863 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
865 struct kioctx_cpu *kcpu;
868 local_irq_save(flags);
869 kcpu = this_cpu_ptr(ctx->cpu);
870 kcpu->reqs_available += nr;
872 while (kcpu->reqs_available >= ctx->req_batch * 2) {
873 kcpu->reqs_available -= ctx->req_batch;
874 atomic_add(ctx->req_batch, &ctx->reqs_available);
877 local_irq_restore(flags);
880 static bool get_reqs_available(struct kioctx *ctx)
882 struct kioctx_cpu *kcpu;
886 local_irq_save(flags);
887 kcpu = this_cpu_ptr(ctx->cpu);
888 if (!kcpu->reqs_available) {
889 int old, avail = atomic_read(&ctx->reqs_available);
892 if (avail < ctx->req_batch)
896 avail = atomic_cmpxchg(&ctx->reqs_available,
897 avail, avail - ctx->req_batch);
898 } while (avail != old);
900 kcpu->reqs_available += ctx->req_batch;
904 kcpu->reqs_available--;
906 local_irq_restore(flags);
910 /* refill_reqs_available
911 * Updates the reqs_available reference counts used for tracking the
912 * number of free slots in the completion ring. This can be called
913 * from aio_complete() (to optimistically update reqs_available) or
914 * from aio_get_req() (the we're out of events case). It must be
915 * called holding ctx->completion_lock.
917 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
920 unsigned events_in_ring, completed;
922 /* Clamp head since userland can write to it. */
923 head %= ctx->nr_events;
925 events_in_ring = tail - head;
927 events_in_ring = ctx->nr_events - (head - tail);
929 completed = ctx->completed_events;
930 if (events_in_ring < completed)
931 completed -= events_in_ring;
938 ctx->completed_events -= completed;
939 put_reqs_available(ctx, completed);
942 /* user_refill_reqs_available
943 * Called to refill reqs_available when aio_get_req() encounters an
944 * out of space in the completion ring.
946 static void user_refill_reqs_available(struct kioctx *ctx)
948 spin_lock_irq(&ctx->completion_lock);
949 if (ctx->completed_events) {
950 struct aio_ring *ring;
953 /* Access of ring->head may race with aio_read_events_ring()
954 * here, but that's okay since whether we read the old version
955 * or the new version, and either will be valid. The important
956 * part is that head cannot pass tail since we prevent
957 * aio_complete() from updating tail by holding
958 * ctx->completion_lock. Even if head is invalid, the check
959 * against ctx->completed_events below will make sure we do the
962 ring = kmap_atomic(ctx->ring_pages[0]);
966 refill_reqs_available(ctx, head, ctx->tail);
969 spin_unlock_irq(&ctx->completion_lock);
973 * Allocate a slot for an aio request.
974 * Returns NULL if no requests are free.
976 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
978 struct aio_kiocb *req;
980 if (!get_reqs_available(ctx)) {
981 user_refill_reqs_available(ctx);
982 if (!get_reqs_available(ctx))
986 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
990 percpu_ref_get(&ctx->reqs);
995 put_reqs_available(ctx, 1);
999 static void kiocb_free(struct aio_kiocb *req)
1001 if (req->common.ki_filp)
1002 fput(req->common.ki_filp);
1003 if (req->ki_eventfd != NULL)
1004 eventfd_ctx_put(req->ki_eventfd);
1005 kmem_cache_free(kiocb_cachep, req);
1008 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1010 struct aio_ring __user *ring = (void __user *)ctx_id;
1011 struct mm_struct *mm = current->mm;
1012 struct kioctx *ctx, *ret = NULL;
1013 struct kioctx_table *table;
1016 if (get_user(id, &ring->id))
1020 table = rcu_dereference(mm->ioctx_table);
1022 if (!table || id >= table->nr)
1025 ctx = table->table[id];
1026 if (ctx && ctx->user_id == ctx_id) {
1027 percpu_ref_get(&ctx->users);
1036 * Called when the io request on the given iocb is complete.
1038 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1040 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1041 struct kioctx *ctx = iocb->ki_ctx;
1042 struct aio_ring *ring;
1043 struct io_event *ev_page, *event;
1044 unsigned tail, pos, head;
1045 unsigned long flags;
1048 * Special case handling for sync iocbs:
1049 * - events go directly into the iocb for fast handling
1050 * - the sync task with the iocb in its stack holds the single iocb
1051 * ref, no other paths have a way to get another ref
1052 * - the sync task helpfully left a reference to itself in the iocb
1054 BUG_ON(is_sync_kiocb(kiocb));
1056 if (iocb->ki_list.next) {
1057 unsigned long flags;
1059 spin_lock_irqsave(&ctx->ctx_lock, flags);
1060 list_del(&iocb->ki_list);
1061 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1065 * Add a completion event to the ring buffer. Must be done holding
1066 * ctx->completion_lock to prevent other code from messing with the tail
1067 * pointer since we might be called from irq context.
1069 spin_lock_irqsave(&ctx->completion_lock, flags);
1072 pos = tail + AIO_EVENTS_OFFSET;
1074 if (++tail >= ctx->nr_events)
1077 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1078 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1080 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1081 event->data = iocb->ki_user_data;
1085 kunmap_atomic(ev_page);
1086 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1088 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1089 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1092 /* after flagging the request as done, we
1093 * must never even look at it again
1095 smp_wmb(); /* make event visible before updating tail */
1099 ring = kmap_atomic(ctx->ring_pages[0]);
1102 kunmap_atomic(ring);
1103 flush_dcache_page(ctx->ring_pages[0]);
1105 ctx->completed_events++;
1106 if (ctx->completed_events > 1)
1107 refill_reqs_available(ctx, head, tail);
1108 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1110 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1113 * Check if the user asked us to deliver the result through an
1114 * eventfd. The eventfd_signal() function is safe to be called
1117 if (iocb->ki_eventfd != NULL)
1118 eventfd_signal(iocb->ki_eventfd, 1);
1120 /* everything turned out well, dispose of the aiocb. */
1124 * We have to order our ring_info tail store above and test
1125 * of the wait list below outside the wait lock. This is
1126 * like in wake_up_bit() where clearing a bit has to be
1127 * ordered with the unlocked test.
1131 if (waitqueue_active(&ctx->wait))
1132 wake_up(&ctx->wait);
1134 percpu_ref_put(&ctx->reqs);
1137 /* aio_read_events_ring
1138 * Pull an event off of the ioctx's event ring. Returns the number of
1141 static long aio_read_events_ring(struct kioctx *ctx,
1142 struct io_event __user *event, long nr)
1144 struct aio_ring *ring;
1145 unsigned head, tail, pos;
1150 * The mutex can block and wake us up and that will cause
1151 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1152 * and repeat. This should be rare enough that it doesn't cause
1153 * peformance issues. See the comment in read_events() for more detail.
1155 sched_annotate_sleep();
1156 mutex_lock(&ctx->ring_lock);
1158 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1159 ring = kmap_atomic(ctx->ring_pages[0]);
1162 kunmap_atomic(ring);
1165 * Ensure that once we've read the current tail pointer, that
1166 * we also see the events that were stored up to the tail.
1170 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1175 head %= ctx->nr_events;
1176 tail %= ctx->nr_events;
1180 struct io_event *ev;
1183 avail = (head <= tail ? tail : ctx->nr_events) - head;
1187 avail = min(avail, nr - ret);
1188 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1189 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1191 pos = head + AIO_EVENTS_OFFSET;
1192 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1193 pos %= AIO_EVENTS_PER_PAGE;
1196 copy_ret = copy_to_user(event + ret, ev + pos,
1197 sizeof(*ev) * avail);
1200 if (unlikely(copy_ret)) {
1207 head %= ctx->nr_events;
1210 ring = kmap_atomic(ctx->ring_pages[0]);
1212 kunmap_atomic(ring);
1213 flush_dcache_page(ctx->ring_pages[0]);
1215 pr_debug("%li h%u t%u\n", ret, head, tail);
1217 mutex_unlock(&ctx->ring_lock);
1222 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1223 struct io_event __user *event, long *i)
1225 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1230 if (unlikely(atomic_read(&ctx->dead)))
1236 return ret < 0 || *i >= min_nr;
1239 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1240 struct io_event __user *event,
1241 struct timespec __user *timeout)
1243 ktime_t until = { .tv64 = KTIME_MAX };
1249 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1252 until = timespec_to_ktime(ts);
1256 * Note that aio_read_events() is being called as the conditional - i.e.
1257 * we're calling it after prepare_to_wait() has set task state to
1258 * TASK_INTERRUPTIBLE.
1260 * But aio_read_events() can block, and if it blocks it's going to flip
1261 * the task state back to TASK_RUNNING.
1263 * This should be ok, provided it doesn't flip the state back to
1264 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1265 * will only happen if the mutex_lock() call blocks, and we then find
1266 * the ringbuffer empty. So in practice we should be ok, but it's
1267 * something to be aware of when touching this code.
1269 if (until.tv64 == 0)
1270 aio_read_events(ctx, min_nr, nr, event, &ret);
1272 wait_event_interruptible_hrtimeout(ctx->wait,
1273 aio_read_events(ctx, min_nr, nr, event, &ret),
1276 if (!ret && signal_pending(current))
1283 * Create an aio_context capable of receiving at least nr_events.
1284 * ctxp must not point to an aio_context that already exists, and
1285 * must be initialized to 0 prior to the call. On successful
1286 * creation of the aio_context, *ctxp is filled in with the resulting
1287 * handle. May fail with -EINVAL if *ctxp is not initialized,
1288 * if the specified nr_events exceeds internal limits. May fail
1289 * with -EAGAIN if the specified nr_events exceeds the user's limit
1290 * of available events. May fail with -ENOMEM if insufficient kernel
1291 * resources are available. May fail with -EFAULT if an invalid
1292 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1295 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1297 struct kioctx *ioctx = NULL;
1301 ret = get_user(ctx, ctxp);
1306 if (unlikely(ctx || nr_events == 0)) {
1307 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1312 ioctx = ioctx_alloc(nr_events);
1313 ret = PTR_ERR(ioctx);
1314 if (!IS_ERR(ioctx)) {
1315 ret = put_user(ioctx->user_id, ctxp);
1317 kill_ioctx(current->mm, ioctx, NULL);
1318 percpu_ref_put(&ioctx->users);
1326 * Destroy the aio_context specified. May cancel any outstanding
1327 * AIOs and block on completion. Will fail with -ENOSYS if not
1328 * implemented. May fail with -EINVAL if the context pointed to
1331 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1333 struct kioctx *ioctx = lookup_ioctx(ctx);
1334 if (likely(NULL != ioctx)) {
1335 struct completion requests_done =
1336 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1339 /* Pass requests_done to kill_ioctx() where it can be set
1340 * in a thread-safe way. If we try to set it here then we have
1341 * a race condition if two io_destroy() called simultaneously.
1343 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1344 percpu_ref_put(&ioctx->users);
1346 /* Wait until all IO for the context are done. Otherwise kernel
1347 * keep using user-space buffers even if user thinks the context
1351 wait_for_completion(&requests_done);
1355 pr_debug("EINVAL: invalid context id\n");
1359 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1360 unsigned long, loff_t);
1361 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1363 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1364 struct iovec **iovec,
1366 struct iov_iter *iter)
1368 #ifdef CONFIG_COMPAT
1370 return compat_import_iovec(rw,
1371 (struct compat_iovec __user *)buf,
1372 len, UIO_FASTIOV, iovec, iter);
1374 return import_iovec(rw, (struct iovec __user *)buf,
1375 len, UIO_FASTIOV, iovec, iter);
1380 * Performs the initial checks and io submission.
1382 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1383 char __user *buf, size_t len, bool compat)
1385 struct file *file = req->ki_filp;
1390 rw_iter_op *iter_op;
1391 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1392 struct iov_iter iter;
1395 case IOCB_CMD_PREAD:
1396 case IOCB_CMD_PREADV:
1399 rw_op = file->f_op->aio_read;
1400 iter_op = file->f_op->read_iter;
1403 case IOCB_CMD_PWRITE:
1404 case IOCB_CMD_PWRITEV:
1407 rw_op = file->f_op->aio_write;
1408 iter_op = file->f_op->write_iter;
1411 if (unlikely(!(file->f_mode & mode)))
1414 if (!rw_op && !iter_op)
1417 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1418 ret = aio_setup_vectored_rw(rw, buf, len,
1419 &iovec, compat, &iter);
1421 ret = import_single_range(rw, buf, len, iovec, &iter);
1425 ret = rw_verify_area(rw, file, &req->ki_pos,
1426 iov_iter_count(&iter));
1434 /* XXX: move/kill - rw_verify_area()? */
1435 /* This matches the pread()/pwrite() logic */
1436 if (req->ki_pos < 0) {
1442 file_start_write(file);
1445 ret = iter_op(req, &iter);
1447 ret = rw_op(req, iter.iov, iter.nr_segs, req->ki_pos);
1451 file_end_write(file);
1455 case IOCB_CMD_FDSYNC:
1456 if (!file->f_op->aio_fsync)
1459 ret = file->f_op->aio_fsync(req, 1);
1462 case IOCB_CMD_FSYNC:
1463 if (!file->f_op->aio_fsync)
1466 ret = file->f_op->aio_fsync(req, 0);
1470 pr_debug("EINVAL: no operation provided\n");
1474 if (ret != -EIOCBQUEUED) {
1476 * There's no easy way to restart the syscall since other AIO's
1477 * may be already running. Just fail this IO with EINTR.
1479 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1480 ret == -ERESTARTNOHAND ||
1481 ret == -ERESTART_RESTARTBLOCK))
1483 aio_complete(req, ret, 0);
1489 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1490 struct iocb *iocb, bool compat)
1492 struct aio_kiocb *req;
1495 /* enforce forwards compatibility on users */
1496 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1497 pr_debug("EINVAL: reserve field set\n");
1501 /* prevent overflows */
1503 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1504 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1505 ((ssize_t)iocb->aio_nbytes < 0)
1507 pr_debug("EINVAL: overflow check\n");
1511 req = aio_get_req(ctx);
1515 req->common.ki_filp = fget(iocb->aio_fildes);
1516 if (unlikely(!req->common.ki_filp)) {
1520 req->common.ki_pos = iocb->aio_offset;
1521 req->common.ki_complete = aio_complete;
1522 req->common.ki_flags = 0;
1524 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1526 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1527 * instance of the file* now. The file descriptor must be
1528 * an eventfd() fd, and will be signaled for each completed
1529 * event using the eventfd_signal() function.
1531 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1532 if (IS_ERR(req->ki_eventfd)) {
1533 ret = PTR_ERR(req->ki_eventfd);
1534 req->ki_eventfd = NULL;
1538 req->common.ki_flags |= IOCB_EVENTFD;
1541 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1542 if (unlikely(ret)) {
1543 pr_debug("EFAULT: aio_key\n");
1547 req->ki_user_iocb = user_iocb;
1548 req->ki_user_data = iocb->aio_data;
1550 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1551 (char __user *)(unsigned long)iocb->aio_buf,
1559 put_reqs_available(ctx, 1);
1560 percpu_ref_put(&ctx->reqs);
1565 long do_io_submit(aio_context_t ctx_id, long nr,
1566 struct iocb __user *__user *iocbpp, bool compat)
1571 struct blk_plug plug;
1573 if (unlikely(nr < 0))
1576 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1577 nr = LONG_MAX/sizeof(*iocbpp);
1579 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1582 ctx = lookup_ioctx(ctx_id);
1583 if (unlikely(!ctx)) {
1584 pr_debug("EINVAL: invalid context id\n");
1588 blk_start_plug(&plug);
1591 * AKPM: should this return a partial result if some of the IOs were
1592 * successfully submitted?
1594 for (i=0; i<nr; i++) {
1595 struct iocb __user *user_iocb;
1598 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1603 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1608 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1612 blk_finish_plug(&plug);
1614 percpu_ref_put(&ctx->users);
1619 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1620 * the number of iocbs queued. May return -EINVAL if the aio_context
1621 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1622 * *iocbpp[0] is not properly initialized, if the operation specified
1623 * is invalid for the file descriptor in the iocb. May fail with
1624 * -EFAULT if any of the data structures point to invalid data. May
1625 * fail with -EBADF if the file descriptor specified in the first
1626 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1627 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1628 * fail with -ENOSYS if not implemented.
1630 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1631 struct iocb __user * __user *, iocbpp)
1633 return do_io_submit(ctx_id, nr, iocbpp, 0);
1637 * Finds a given iocb for cancellation.
1639 static struct aio_kiocb *
1640 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1642 struct aio_kiocb *kiocb;
1644 assert_spin_locked(&ctx->ctx_lock);
1646 if (key != KIOCB_KEY)
1649 /* TODO: use a hash or array, this sucks. */
1650 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1651 if (kiocb->ki_user_iocb == iocb)
1658 * Attempts to cancel an iocb previously passed to io_submit. If
1659 * the operation is successfully cancelled, the resulting event is
1660 * copied into the memory pointed to by result without being placed
1661 * into the completion queue and 0 is returned. May fail with
1662 * -EFAULT if any of the data structures pointed to are invalid.
1663 * May fail with -EINVAL if aio_context specified by ctx_id is
1664 * invalid. May fail with -EAGAIN if the iocb specified was not
1665 * cancelled. Will fail with -ENOSYS if not implemented.
1667 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1668 struct io_event __user *, result)
1671 struct aio_kiocb *kiocb;
1675 ret = get_user(key, &iocb->aio_key);
1679 ctx = lookup_ioctx(ctx_id);
1683 spin_lock_irq(&ctx->ctx_lock);
1685 kiocb = lookup_kiocb(ctx, iocb, key);
1687 ret = kiocb_cancel(kiocb);
1691 spin_unlock_irq(&ctx->ctx_lock);
1695 * The result argument is no longer used - the io_event is
1696 * always delivered via the ring buffer. -EINPROGRESS indicates
1697 * cancellation is progress:
1702 percpu_ref_put(&ctx->users);
1708 * Attempts to read at least min_nr events and up to nr events from
1709 * the completion queue for the aio_context specified by ctx_id. If
1710 * it succeeds, the number of read events is returned. May fail with
1711 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1712 * out of range, if timeout is out of range. May fail with -EFAULT
1713 * if any of the memory specified is invalid. May return 0 or
1714 * < min_nr if the timeout specified by timeout has elapsed
1715 * before sufficient events are available, where timeout == NULL
1716 * specifies an infinite timeout. Note that the timeout pointed to by
1717 * timeout is relative. Will fail with -ENOSYS if not implemented.
1719 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1722 struct io_event __user *, events,
1723 struct timespec __user *, timeout)
1725 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1728 if (likely(ioctx)) {
1729 if (likely(min_nr <= nr && min_nr >= 0))
1730 ret = read_events(ioctx, min_nr, nr, events, timeout);
1731 percpu_ref_put(&ioctx->users);